How to Generate Videos with FLUX 3 (2026)

FLUX 3 is Black Forest Labs' multimodal model, announced on July 23, 2026, and one of the first big AI video releases where dialogue, effects, ambience and picture come out of the same generation. Most FLUX 3 guides still repeat the launch announcement, and a clip that costs 170 credits at its shortest needs a prompt that works the first time. So we read all 13 FLUX 3 clips on the public canvases of Flick team and residency films, with the exact prompts that made them and the sound tracks that came back. This guide shows how to generate videos with FLUX 3, the real prompts, how to direct the sound, what it costs and how it compares with Seedance, Kling, Veo and Grok.

FLUX 3 runs on Flick's Standard plan and up.

Can FLUX 3 generate videos?

What is FLUX 3?

Black Forest Labs announced FLUX 3 on July 23, 2026 and made it generally available through its API and partners on August 4, 2026.

BFL CEO Robin Rombach summed up the premise: "You can't cheat reality. A model that only learns images can only generate images."

That framing matters to filmmakers, because the hard part of AI video is no longer a single pretty frame but continuity: picture, motion, sound, character, physics and edit rhythm all have to agree.

• Read the benchmark as a vendor claim: independent tests have not caught up with it yet.
• The practical change is sound: FLUX 3 makes video and audio in one generation, which changes how a film is put together.
• Early tests: testers posted their first clips on X, such as @venturetwins and @umesh_ai, and the right test is whether motion, sound, timing and scene logic hold together, not still-frame beauty.

What do real FLUX 3 prompts look like?

These are FLUX 3 shots from Flick team and residency films, each with the prompt exactly as it was run.

• What we read: all 13 FLUX 3 clips on the public canvases of 4 Flick team and residency films, 7 of them at the full 20 seconds.
• Sound in every file: all 13 clips came back with a sound track, including the 7 whose prompts never mention sound.
• Start frames: all 13 began from a start frame, so we have no FLUX 3 clip of our own made from text alone yet.

00:10

A locked wide shot with every sound on its own line

The prompt locks the camera, then lists the wind, the hooves, the tack and the horse's breath line by line and rules out dialogue and music, so the sound is written as carefully as the picture.

Locked wide shot.
Maintain the same camera position, composition, focal length, and shot scale throughout.
No push-in, pull-back, pan, tilt, crane movement, or zoom.
The rider and horse move independently from the camera.

The rider and horse maintain a fast gallop along their existing direction across the frozen lake toward the distant tree line.

The horse sustains a natural, powerful gallop. All four legs move through a complete running gait, while its head and neck rise and fall naturally with each stride. Its mane, tail, and tack move with the speed.

The rider leans slightly forward, keeps both hands steady on the reins, and moves naturally with the horse’s galloping rhythm.

The hooves repeatedly strike the thin snow covering the ice, throwing a small amount of fine snow backward and extending faint hoof tracks behind them.

The colossal serpent in the background is a completely frozen corpse and remains absolutely motionless throughout. No breathing, blinking, eye movement, jaw movement, or body movement.

The broad background fog moves slowly throughout the shot, drifting sideways at a steady, continuous pace across the mountain and around the frozen serpent. The fog banks change shape naturally. The moving fog does not affect the serpent, its ice covering, or its icicles; the serpent remains absolutely motionless.

Continuous low cold wind.
Fast, forceful hoof impacts against the ice and thin snow.
Subtle rein, leather, and tack vibration.
Heavy, rhythmic breathing from the running horse.
No sound from the serpent.
No dialogue.
No background music.

Jaime Yao, director of In ’twixt Vows

00:06

A slow detail shot that says what must stay still

The prompt names the few tiles that may move, how slowly, and what must not, down to the figure not turning or blinking, which keeps a slow detail shot from becoming a collapse.

The camera drifts extremely slowly from upper right to lower left with a very slight push-in. No sudden acceleration. No reframing.

The mosaic figure remains still. Only a sparse selection of mosaic pieces slowly loosens, lifts away from the wall, and floats just in front of their original positions. Do not make it a large-scale collapse.

After the pieces detach, the exposed gaps reveal a dim blue moonlit jade material. It is smooth, semi-translucent, faintly glowing, and not liquid. It emits a soft, low-intensity cool blue glow that gently illuminates nearby piece edges, seams, and fine airborne dust.

A few pieces near the blue gem eye may also loosen and float. The gem catches one brief subtle glint. The figure does not turn its head and does not blink.

Only a small amount of dust and faint haze drifts slowly through the air.

Jaime Yao, director of In ’twixt Vows

00:10

Four words on a strange start frame

With four words and a typo, FLUX 3 kept the start frame's floating car and its two watchers and set the car moving, its boot opening and smoke rising as it sinks toward the sea.

some strange things happned.

Zoey Z., director of Summer time

00:20

Six words that leave the shot to the model

A prompt this open leaves every choice to FLUX 3, and this take moves from the bedroom start frame to a second scene of a woman in a trench coat beside a red car.

based on these images, create something

Zoey Z., director of Summer time

How do you direct the sound in FLUX 3?

In The Miracle, Sam Wasserman wrote an off-camera line for his General, described the voice in detail and ruled out music and other voices, and FLUX 3 returned the picture and the voice in one pass:

00:20

Dense ranks of mixed male and female Caucasian Runners with varied natural hair colors and clearly distinct human faces wear the same Pearl-White Gauntlet Suit: Sleek, slightly glossy pearl-white suit over an open-collar white shirt; identical material across all Runners while faces, ages, genders, builds, and hair remain clearly distinct. The General remains completely off camera. His exact voice continues: “It will push you to your breaking point, but nothing can prepare you for the true brutality of the Gauntlet.” Match the supplied General voice reference exactly: low grounded baritone, natural accent, firm chest resonance, hard consonants, dry weathered texture, restrained authority and controlled grief.
# Composition
Track laterally past tense hands, breathing chests, worn faces and fixed eyes, revealing multiple Runners in layered depth. No one speaks or lip-syncs.
# Lighting
Sodium-amber practicals, muted cyan shadows, smoke, wet-floor reflections and deep blacks.
# Camera
20-second medium shot, eye level, 50mm anamorphic, slow weighty lateral track. Let the voice finish cleanly before the final reaction.
# Style
Lost large-scale 1987 live-action 35mm anamorphic cinema; coarse organic grain, gate weave, halation, practical atmosphere, believable human skin and physical costumes. No synthetic faces, CGI sheen, subtitles, readable incidental text, music, or additional voices.
# Mood
Quiet dread held under military discipline.

• The voice: the track carries speech for roughly the first 12 seconds, then room sound with no voice to the end, as the prompt's "Let the voice finish cleanly before the final reaction" asks.
• No music: the prompt rules out music and additional voices, and the track holds none after the line ends.
• No voice reference in Flick: the prompt asks to match a "supplied General voice reference", but FLUX 3 in Flick takes a start and an end frame and no audio input, so the voice comes from the written description alone.
• The same prompt twice: Sam ran this prompt from two different start frames, and the second take opens this guide.

Before synchronized AI video, the usual workflow was to generate a silent clip, pick the best take, add dialogue with a voice model, add foley and ambience from a library, add music, then try to make it all feel like one room.

FLUX 3 folds some of that into the generation: a door can slam with the sound of the door, an alarm can pulse with the red light, and a character can speak in the shot where the camera moves.

That does not make post-production irrelevant, it makes the first pass more film-like, since you still edit, mix and grade, but the raw clip may arrive with its timing and sound cues attached.

How do you generate a video with FLUX 3 in Flick?

1. 1Open a canvas

   Create a Flick project in the aspect ratio of the final release, 16:9, 9:16 or 1:1, and keep prompt notes, scene stills and references beside the shot plan.

2. 2Lock characters first

   Anchor recurring characters in stills with Character Reference before you rely on a video model, because every clip can drift from shot to shot.

3. 3Pick FLUX 3

   Add an Image to Video node, choose FLUX 3, and add a start frame, or a start and an end frame, when the shot must begin or land on a set picture.

4. 4Write shot and sound

   Describe one camera move, the action in order and the sound you want, then set 5 to 20 seconds and 720p or 1080p and check the credit total.

5. 5Compare models

   Use FLUX for sound-aware shots, Seedance for reference-heavy or longer clips, Kling for strong motion and Veo for polished realism, while the film stays on one canvas.

6. 6Extract, grade, edit

   Extract a strong final frame as the next shot's start, then grade and cut FLUX clips, silent clips, voices and music together so the sequence reads as one film.

How do you write a good FLUX 3 prompt?

1. 1. Start with a shot, not a vague idea

   Say who is in frame, where the camera is, what moves, what changes over time, what the audio holds and what the model should avoid, because FLUX 3 still needs a shot to follow.

2. 2. Write picture and sound together

   Give the sound its own line, as the frozen lake prompt does with wind, hooves and breathing, since FLUX 3 makes the audio from the same prompt as the picture.

3. 3. Keep the first test simple

   Use one subject, one camera idea and one sound idea for early generations, and leave scenes that need heavy choreography until you know how the model behaves.

4. 4. Say what must stay still

   Name what must not move as clearly as what should, as the mosaic prompt does with the figure that does not turn or blink, so a slow shot keeps its shape.

5. 5. Compare takes by how they cut

   Ask which clip cuts best rather than which looks best, checking that the first and last frames meet the shots around it and that the sound helps the edit.

6. 6. Save the full length for shots that earn it

   Keep cutaways, inserts and reaction shots short and save the longest clips for a stable subject with a beginning, middle and end, because AI video often looks strongest when a shot has one purpose.

A strong FLUX 3 prompt covers these six things:

• Camera: push-in, dolly, handheld, locked-off, aerial, tracking, orbit.
• Subject motion: walks, turns, reaches, runs, hesitates, looks up.
• Environmental motion: rain, smoke, wind, dust, crowds, fire, water.
• Sound: footsteps, cloth, engine hum, thunder, dialogue, ambience, silence.
• Timing: "after a beat," "as the camera reaches them," "at the end."
• Constraints: no text, no extra characters, preserve identity, realistic motion.

A weak prompt

A futuristic city at night.

A better prompt, with picture and sound

Slow aerial push over a rain-soaked futuristic city at night. Neon signs reflect in puddles below, steam rises from street vents, distant traffic hums, and a low synth drone builds under the shot. Cinematic realism, wet asphalt, volumetric light, no readable text, no logos.

When you review takes, check that the first frame matches the scene before it, the last frame gives a cut point, the action reads without explanation, the subject keeps its identity and the motion feels physically plausible.

Is there a FLUX 3 prompt template?

[Shot type / camera movement] of [subject] in [setting].
[Subject action over time].
[Environmental motion over time].
Audio: [dialogue, ambience, effects, silence, music direction].
Style: [cinematic style, lens, lighting, texture].
Constraints: [identity, no text, no logos, no extra characters, realistic motion].

The template filled in

Slow handheld tracking shot behind a detective walking down a narrow motel hallway at midnight.
The detective moves cautiously, one hand near the wall, then stops when a door creaks open at the end of the hall.
Fluorescent lights flicker overhead, dust floats in the air, rain runs down the window at the far end.
Audio: soft footsteps on carpet, low motel electricity hum, distant rain, one sharp door creak, no music.
1970s neo-noir, 35mm film grain, warm green fluorescent cast, realistic motion.
No readable text, no logos, no extra people.

Locked-off medium shot of a tired astronaut sitting alone inside a dim spacecraft cockpit.
At the start, only the soft instrument panel glow lights their face. After three seconds, a red warning light begins pulsing.
The astronaut slowly turns toward the window as distant debris taps against the hull.
Audio: low spacecraft hum, faint alarm pulse, subtle breathing inside the helmet, no music.
Cinematic realism, shallow depth of field, no text, no extra characters.

Close-up of an exhausted pilot in a dim cockpit, helmet visor half raised.
The pilot looks down, breathes once, then says quietly: "We are not going back."
Warning lights pulse across their face as the camera slowly pushes in.
Audio: cockpit hum, soft alarm beeps, intimate dry dialogue, no music.
Cinematic sci-fi realism, shallow depth of field, preserve face, no text.

Locked-off wide shot of an empty farmhouse hallway at night.
At first nothing moves. After four seconds, a floorboard creaks off-screen, then the hanging light swings slightly.
Audio: deep room tone, distant wind, single floorboard creak, faint chain movement, no music.
Natural darkness, practical bulb light, realistic horror atmosphere, no visible monster.

Slow studio orbit around a matte black cinema camera on a reflective table.
Light sweeps across the lens glass as the camera rotates, revealing subtle dust particles in the air.
Audio: quiet mechanical rotation, soft room tone, no music.
Premium commercial lighting, crisp reflections, minimal background.
No logos, no text, keep object shape consistent.

Wide cinematic shot of ancient ships cutting through dark water at dawn.
Oars move in rhythm, sails snap in the wind, mist wraps around rocky cliffs.
Audio: waves against wood, distant sailcloth snapping, low wind, subtle drum pulse.
Historical epic realism, warm sunrise, 70mm texture, grounded scale.
No modern objects, no fantasy creatures, no readable text.

Medium close-up of a young inventor sitting alone at a cluttered workbench before sunrise.
They tighten one small screw, stop, and smile when the machine begins to hum.
Dust floats in the first blue morning light through the window.
Audio: tiny metal screw turn, soft electric hum, morning birds outside, no music.
Naturalistic indie film style, gentle lens softness, realistic hands.

How much does FLUX 3 cost in Flick?

• Sound is included: FLUX 3 in Flick has no audio switch, so the price covers picture and sound together.
• Plans: FLUX 3 needs the Standard plan, from $20 a month with 2,500 credits.
• Test short first: the node shows the credit total before you run, so try a prompt at the shortest length before a full-length take.

How does FLUX 3 compare with other AI video models?

• References: Seedance 2.5 in Reference to Video takes up to 30 reference images, while FLUX 3 in Flick takes a start frame and an end frame and no reference images.
• The canvas is the constant: Character Reference, prompt notes and the edit stay on one Flick canvas, so a film can use a different model for each shot.
• A head to head test: FLUX 3 vs Seedance 2.5 runs both models on the same shots.

What can FLUX 3 not do yet?

• Directing the audio: our clips show that written sound lines and a quoted spoken line come through, but how precisely a prompt controls dialogue, foley, ambience and music still needs more tests.
• The same performance twice: filmmakers need to regenerate the picture while keeping a performance, and FLUX 3 offers no way to do that yet.
• A voice across shots: native sound matters most if a character sounds the same from shot to shot, and FLUX 3 in Flick takes no voice reference.
• References: character, wardrobe and location consistency decide whether it works for real films, and FLUX 3 in Flick takes no reference images, so lock looks in the start frame.
• Cost of iteration: the credit cost shows before you generate, so test with short clips first.
• Open weights: the developer versions Black Forest Labs promised could matter more than the hosted release in the long run.

What are common FLUX 3 mistakes?

• Ignoring audio in the prompt: FLUX 3's difference is synchronized sound, so direct it.
• A full scene in one generation: one clip should be one shot.
• Overusing the full length: a longer clip is better when the action stays coherent, and worse when it does not.
• Judging the image alone: listen to timing, sound design, ambience and performance.
• Taking vendor benchmarks as final: independent comparisons will matter.
• Forgetting continuity: a beautiful clip still has to match the character, location and sequence.
• Waiting for one perfect model: use the best model per shot and assemble the film in one workflow.

FLUX 3 FAQ

More on FLUX 3 and AI video in Flick

Films in this guide

• Sam Wasserman · The Miracle
• Jaime Yao · In ’twixt Vows
• Zoey Z. · Summer time
• Raj Rishi More · Seasons
• MaxVOAO · Dreamland
• Owen Doerr · Lifeguard